{"id":"edb52b6e-2dec-4eb7-890d-31dbb7752ebd","arxiv_id":"2412.20314","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A BCD-based algorithm jointly allocates time, frequency, power, and beam resources across frames to minimize target-tracking PCRB while maintaining UE throughput in an ISAC network.","lead":"This paper designs an algorithm to split scarce radio resources between sensing and communication in 6G networks, so that targets can be tracked in every frame while users still get their data rates. The authors formulate the joint allocation problem, solve it with block coordinate descent, and show in simulation that their scheme beats a random baseline and approaches an optimistic upper bound.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Measurement-model premise in (11)-(12) is under-specified: proportionality constants are omitted and beam-prediction mismatch is not modeled, so the optimized PCRB may not correspond to actual tracking error.","rationale":"After reading the full manuscript, the most load-bearing condition for the central claim is the measurement model in (11)-(15). The paper's objective is a PCRB built on these CRB expressions, and the EKF update in (18) uses the same covariance as the true measurement noise. Three specific gaps make this premise insecure: (i) the proportionality in (11)-(12) omits constants, and because (15) mixes range and angle info via Q, an unknown ratio between these constants changes the optimized allocation; (ii) the beamformer in (4) is set to the predicted AoD, but the CRB uses ||H_m f_m||^2 with no beam-pointing-loss factor, so under the dynamicity the paper emphasizes, the predicted SNR is optimistic; (iii) the CRB is a lower bound, and no estimator achieving it is identified or simulated, so the EKF's covariance in (18) may be overconfident. The reader identified this as the weakest assumption; I agree. The integrality issue in P1.4 (integer constraints present while the problem is claimed convex) is a real implementation gap, but it can be addressed by rounding or a MINLP solver without changing the framework. Similarly, the infeasible upper bound and weak baseline weaken the empirical evidence but do not invalidate the approach. Thus the measurement model is the single most load-bearing concern. A concrete derivation or estimator simulation would settle it.","tokens_in":10688,"tokens_out":16529,"duration_ms":161480,"concrete_test":"Derive the exact range/AoA CRBs from the signal model (2) (including constants from [18] and the beam gain |a_t^H(θ_m)a_t(θhat_m)|^2), or simulate the actual estimators (e.g., FFT-based range and monopulse angle) and compare the empirical variance to (11)-(12). Then rerun Algorithm 1 with the corrected Σ_m and check the resulting PCRB and QoS feasibility in Fig. 3. If the optimal allocation changes or the reported sensing gain over RFTEP drops by more than 10%, the model premise is load-bearing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Equations (11) and (12) give the distance and AoA error variances only up to proportionality, with no constants; these variances enter the data FIM via (15) and determine the relative weight of range vs. angle information in the objective (25a). If the true constants differ between range and angle (or are not both 1), the optimized allocation and the reported 28.57%/53.95% gains are not reproducible. Independently, the beamformer f_m is fixed to the predicted AoD in (4), but the CRBs in (11)-(12) use ||H_m f_m||^2 without the beam gain |a_t^H(θ_m)a_t(θhat_m)|^2. Under the frame-to-frame target movement modeled in (5)-(7), the prediction error is nonzero, so the actual SNR is lower than assumed; the PCRB optimized by Algorithm 1 is optimistic, and the EKF in (16)-(20) is run with an overconfident measurement covariance. The central claim that per-frame scheduling can be done without instant target CSI therefore rests on an unvalidated equivalence between the CRB expressions and the true estimator/beam-matched performance.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper considers a millimeter-wave ISAC network in which a base station tracks M moving targets and serves K single-antenna UEs over multiple frames. The authors formulate a per-frame joint resource allocation problem (P1) that minimizes the trace of the posterior Cramér-Rao bound (PCRB) for target tracking, subject to UE throughput constraints, power constraints, and frequency/time/spatial resource constraints. To solve the non-convex mixed-integer problem, they propose a block coordinate descent (BCD) algorithm that alternates between a block {n_m, o_ki, p_k} handled by penalty, successive convex approximation (SCA), and CVX, and a block {p_m, f_k} handled by a one-dimensional beam search and CVX. Numerical simulations show that the proposed solution tracks all targets, approaches an upper bound on the PCRB in most frames, maintains high QoS feasibility, and outperforms a random-allocation benchmark by 28.57% in sensing and 53.95% in communication. The paper also demonstrates a sensing-communication trade-off as the throughput threshold increases.","tokens_in":10919,"tokens_out":6693,"duration_ms":67896,"significance":"If the claims hold, the paper offers a practical per-frame resource allocation framework for ISAC target tracking under target motion and time-varying channels, without requiring instantaneous target CSI. The core idea of transforming the PCRB objective into an algebraic form via eigendecomposition is sound and follows the established framework of [8]. The paper provides a clearly specified system model, a reproducible simulation setup, and a quantitative comparison against a benchmark. The significance is moderated, however, by two omitted derivations, an under-specified measurement model with proportionality constants absent, and the absence of a direct comparison between the optimized PCRB and the actual tracking MSE. These gaps are fixable but are load-bearing for the numerical claims.","major_comments":[{"comment":"The distance and AoA error variances are given only up to proportionality, with the constants omitted. These variances define the measurement covariance Σ^u_m in Eq. (10), enter the data FIM in Eq. (15), and therefore determine the entire PCRB objective in Eq. (25a). Since the two quantities have different physical dimensions (meters squared and radians squared), the trace in Eq. (25a) is not even dimensionally meaningful unless the missing constants supply the appropriate scaling. As written, the numerical results and the reported 28.57% and 53.95% gains are not reproducible. Please provide the explicit CRB expressions with all constants, or state exactly which constants were used in the simulations.","section":"II-B3, Eqs. (11)-(12)"},{"comment":"The convexity of f(o_ki, p_k) is load-bearing for the claim that P1.4 is convex and hence solvable by CVX, but the proof is omitted with the sentence 'The detailed proof is omitted due to limited space.' Similarly, the derivation of the algebraic form for the sub-problem P2.1 is omitted with 'Detailed derivation is omitted due to limited space.' These are not merely presentational omissions: the algorithm's convergence and the validity of the SCA penalty step depend on these derivations. Please include the Hessian computation for f(o_ki, p_k) and the full derivation for P2.1, or provide a precise reference where each result is proven.","section":"IV, after Eq. (34) and after Eq. (38)"},{"comment":"The PCRB objective is evaluated at the predicted target state, because the Jacobian Q̂^u_m and measurement covariance Σ̂^u_m in Eq. (15) use the predicted state ξ̂^u_m from the EKF (Eqs. (16)-(20)), and the beamformer in Eq. (4) is matched to the predicted AoD. The expressions in Eqs. (11)-(12) do include the beamforming gain through ||H^u_m f^u_m||^2, so the mismatch between the predicted and true AoD is partially captured; however, the prediction error itself is not modeled or bounded. If the EKF prediction error is significant, the optimized PCRB is optimistic and the EKF update in Eq. (18) uses an overconfident measurement covariance. Please state the assumption on the prediction error magnitude and, ideally, compare the simulated tracking MSE with the optimized PCRB to validate that the minimized quantity reflects the actual tracking performance.","section":"II-B3, II-B5, and IV"}],"minor_comments":[{"comment":"The noise term appears as '\\tilde w^u_m' in some places and as 'ewu_m' in the text immediately after Eq. (8); this typo should be corrected.","section":"Eq. (8)"},{"comment":"The initial resource constraint (1) states n^u_m ∈ {1, ..., N_RB-1}, while the optimization constraint (25e) uses n_m ∈ {n_req, n_req+1, ..., N_RB}; the relationship between these two formulations should be clarified.","section":"Eqs. (1) and (25e)"},{"comment":"The 'upper bound' is not precisely defined. If it means allocating the maximum resources to every target and UE simultaneously, that allocation violates constraint (25d); if it means a per-target bound with all resources dedicated to sensing, then it is not an upper bound on the constrained problem. Please define the upper bound explicitly.","section":"Section V, Fig. 3"},{"comment":"The acronym RFTEP is used without expansion; the text later explains that it corresponds to random resource block and mini-slot allocation with equal power and one-dimensional beam search, but this should be stated at first use.","section":"Section V"},{"comment":"The null-to-null beamwidth ϖ_nn is given as 4.14° in Table I, but Eq. (12) uses ϖ_nn in a formula that otherwise mixes units; please state whether ϖ_nn is used in radians or degrees and ensure dimensional consistency.","section":"Eq. (12) and Table I"},{"comment":"Algorithm 1 is titled 'BCD-based algorithm for solving problem P1.1', but it actually solves the per-frame version of the original problem P1; the naming should be made consistent with the problem labels in Section IV.","section":"Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a concise treatment of an interesting problem, but it currently has the feel of a conference paper extended to journal format: two key derivations are omitted, and the measurement model in Eqs. (11)-(12) is incompletely specified. The central algorithmic idea is plausible and the simulations are internally consistent, but the missing constants and omitted proofs must be supplied before the numerical claims can be considered reproducible. The paper would also benefit from a clearer statement of how it differs from the framework of [8], since it relies heavily on that work for the PCRB recursion and the algebraic objective transformation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this is a genuinely incremental but useful ISAC resource allocation paper. The new bit is per-frame joint allocation with EKF prediction so you don't need instant target CSI; prior work [16] only tracks at frame start. That problem is worth stating, and the BCD-SCA-penalty solver is a reasonable way to attack it. The simulation story is internally consistent, with a clear sensing-communication trade-off and no parameter fitting to the results.\n\nI agree with the reader's conditional verdict. The main soft spot is the measurement model in (11)-(12). The variance expressions are given only up to proportionality, with constants omitted. Since those variances enter the data FIM and set the relative weight of range vs. angle, different (unstated) constants change the optimal allocation and the reported 28.57%/53.95% gains. The stress-test note also worries about beam-prediction mismatch, but that part is actually covered: ||H_m f_m||^2 includes the inner product between the true steering vector and the beamformer, so the mismatch is in the norm. What is not covered is the assumption that real estimators achieve these CRBs; if they don't, the EKF runs with an overconfident covariance and the PCRB optimized is the wrong quantity. That's a real concern, though a common one in this literature.\n\nThe omitted derivations are the second issue. The Hessian convexity proof for (34) and the P2.1 transformation are both skipped with 'limited space.' They are likely correct — the forms are standard — but they are load-bearing, and in a journal version they need to appear or be cited properly. The baseline is also weak: random allocation with equal power is a low bar, and the 'upper bound' uses maximum resources for every target and UE simultaneously, which is infeasible rather than an actual bound.\n\nNone of this is fatal. The problem formulation is clean, the algorithm is sensible, and the simulations support the qualitative trade-off. But the numerical gains should be treated as indicative until the measurement constants are specified and the derivations are filled in. I'd send this to peer review with a request for those additions and a stronger baseline. Readers working on ISAC resource allocation will find the per-frame formulation useful.","headline":"Competent, incremental ISAC resource allocation paper whose per-frame EKF formulation is new but whose measurement model and omitted derivations keep the numerical gains from being fully verified.","tokens_in":11447,"tokens_out":2580,"would_cite":false,"duration_ms":25374,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A block coordinate descent solver turns per-frame ISAC resource allocation into convex subproblems, enabling joint target tracking and communication every frame.","keywords":["ISAC","resource allocation","target tracking","posterior Cramér-Rao bound","block coordinate descent","successive convex approximation","beamforming","6G"],"falsifier":"Run the proposed allocation with realistic estimators whose empirical error covariance differs from the CRB formulas in (11)–(12), then compare realized tracking mean squared error with the predicted trace of PCRB; if the optimized allocation does not reduce actual MSE as predicted, the minimized objective is the wrong one.","tokens_in":10457,"feed_emoji":"📡","tokens_out":9219,"duration_ms":85621,"temperature":0.7,"pith_summary":"This paper tackles a practical gap in integrated sensing and communication (ISAC) networks: scheduling both target tracking and user communication in every frame while targets move and channels change. It formulates a joint allocation of frequency resource blocks, time mini-slots, transmit power, and beams that minimizes the trace of the posterior Cramér-Rao bound (PCRB) on tracking error under per-user throughput and resource limits. Because the resulting problem is non-convex with coupled integer and continuous variables, the paper develops a block coordinate descent algorithm combining a penalty method, successive convex approximation, and one-dimensional beam search. Simulations show the algorithm tracks ten moving targets, reaches the resource-rich upper bound in most frames, and outperforms a random allocation benchmark by 28.57% in sensing and 53.95% in communication, exposing the sensing–communication trade-off.","feed_headline":"Per-frame ISAC resource planning solved as convex subproblems","feed_subtitle":"A BCD solver with penalty terms and convex approximation tracks moving targets while meeting communication QoS every frame.","key_machinery":"The central object is the posterior Fisher information matrix $J(\\xi^u_m)=J_p+J_d$, whose inverse is the PCRB used as the objective. The load-bearing manipulation is the decomposition $J(\\xi_m)=E_m+n_m^2 V_m$ (and its power analogue $J_p+p_m\\bar V_m$), whose eigendecomposition reduces the matrix objective to a sum of reciprocals $(a_{mj}+b_{mj}n_m^2)^{-1}$, making the non-convex problem amenable to successive convex approximation. Around this, the solver applies block coordinate descent to separate the variables, a penalty term with an iteratively growing weight to softly enforce the throughput constraint, and a one-dimensional search over a beam codebook for each user.","core_discovery":"The paper's contention is that per-frame ISAC scheduling under dynamicity does not require instant target channel state information. Using an extended Kalman filter to predict target states, it rewrites the PCRB objective at each frame into a sum of scalar terms $\\sum_j (a_{mj}+b_{mj}(n_m^2))^{-1}$ via an eigendecomposition, which turns the bandwidth allocation part into a tractable difference-of-convex form. The block coordinate descent then alternates between a convex subproblem for bandwidth, time, and user power, and a convex subproblem for target power plus a codebook beam search for users. On its own simulation evidence, the trace of PCRB tracks the upper bound in most frames while the QoS feasibility ratio stays near 100%.","pith_inferences":["The BCD decomposition does not depend on the specific CRB proportionality constants; if experiments show the assumed inverse-power CRB scaling is inaccurate, the same solver could be rerun with empirical covariances, though the claimed near-optimality would have to be rechecked.","The same algebraic manipulation works for other sensing objectives, such as worst-target PCRB or determinant of the PCRB, because it only requires the Fisher information matrix to be affine in squared bandwidth and in power.","With larger beam codebooks the one-dimensional search would dominate cost, so gradient-based or hierarchical beam selection is a natural extension for massive MIMO arrays.","The formulation assumes a single line-of-sight base station; multiple base stations or multipath would introduce cross-target Fisher information terms that break the per-target decomposition, pointing toward distributed or consensus-based extensions."],"forward_implications":["Per-frame joint allocation becomes computationally feasible: each BCD iteration solves convex subproblems, so scheduling can be updated every frame as targets and channels evolve.","The PCRB-based formulation gives a policy that reaches the resource-rich upper bound in most simulated frames, meaning the allocation algorithm itself contributes little tracking loss relative to maximum-resource provisioning.","Raising the communication throughput requirement degrades sensing PCRB in a smooth, predictable way, making the sensing–communication trade-off a tunable design knob.","No instant target CSI is required; the extended Kalman filter prediction supplies the sensing beam direction, so the scheme is applicable when target channel estimates are unavailable or outdated."],"supporting_citations":[{"why":"It supplies the PCRB/FIM construction, the Jacobian-based data FIM, and the eigendecomposition trick that reduces the matrix objective to scalar reciprocals.","marker":"[8]"},{"why":"It provides the time-correlated channel model and the dynamicity-aware frequency and power allocation baseline that this paper extends to per-frame joint scheduling.","marker":"[16]"},{"why":"It gives the CRB formulas for distance and AoA, the constant-velocity state transition model, and the multi-target tracking formulation.","marker":"[18]"},{"why":"It contributes the successive convex approximation and penalty-method optimization machinery used inside the BCD solver.","marker":"[19]"},{"why":"It fixes the NR numerology that defines resource blocks and mini-slots, setting the granularity of the time-frequency allocations.","marker":"[20]"}],"fun_headline_variants":["ISAC tracking with predicted targets via BCD and convex approximation","Per-frame ISAC resource planning: BCD with penalty and SCA","EKF-predicted states enable convex per-frame ISAC allocation","Joint time, frequency, power, space optimized for ISAC tracking","BCD algorithm maintains near-100% QoS while tracking targets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole optimization minimizes a posterior error bound built from assumed Cramér-Rao formulas for distance and angle estimation error, so those formulas must match the real estimators actually producing the measurements.","fun_headline_variants_meta":{"raw":{"variants":["ISAC tracking with predicted targets via BCD and convex approximation","Per-frame ISAC resource planning: BCD with penalty and SCA","EKF-predicted states enable convex per-frame ISAC allocation","Joint time, frequency, power, space optimized for ISAC tracking","BCD algorithm maintains near-100% QoS while tracking targets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000406,"raw_usage":{"total_tokens":2081,"prompt_tokens":883,"completion_tokens":1198,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":499,"completion_tokens_details":{"reasoning_tokens":1109}},"tokens_in":499,"tokens_out":1198,"duration_ms":12580,"temperature":1.0,"reasoning_tokens":1109,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:24:04.108308+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the proposed allocation with realistic estimators whose empirical error covariance differs from the CRB formulas in (11)–(12), then compare realized tracking mean squared error with the predicted trace of PCRB; if the optimized allocation does not reduce actual MSE as predicted, the minimized objective is the wrong one.","supporting_citations":[{"cited_title":"Sensing as a service in 6G perceptive networks: A unified framework for ISAC resource allocation,","cited_arxiv_id":null,"evidence_quote":"It supplies the PCRB/FIM construction, the Jacobian-based data FIM, and the eigendecomposition trick that reduces the matrix objective to scalar reciprocals."},{"cited_title":"Impact of channel aging on dual-function radar-communication systems: Performance analysis and resource allocation,","cited_arxiv_id":null,"evidence_quote":"It provides the time-correlated channel model and the dynamicity-aware frequency and power allocation baseline that this paper extends to per-frame joint scheduling."},{"cited_title":"Simultaneous multibeam resource allocation scheme for multiple target tracking,","cited_arxiv_id":null,"evidence_quote":"It gives the CRB formulas for distance and AoA, the constant-velocity state transition model, and the multi-target tracking formulation."},{"cited_title":"Radiorchestra: Proactive management of millimeter-wave self-backhauled small cells via joint optimization of beamforming, user association, rate selection, and admission control,","cited_arxiv_id":null,"evidence_quote":"It contributes the successive convex approximation and penalty-method optimization machinery used inside the BCD solver."},{"cited_title":"3GPP TS 38.211 v17.1.0 5G NR physical channels and modulation,","cited_arxiv_id":null,"evidence_quote":"It fixes the NR numerology that defines resource blocks and mini-slots, setting the granularity of the time-frequency allocations."}],"review_version":1}